What Is Generative AI? How It Works, Examples and How Businesses Use It

What is generative AI? Generative AI (gen AI) is AI that creates new content, such as text, images, code, audio and video, from instructions written in plain language. It learns patterns from huge amounts of data and produces the most likely output for your request. ChatGPT, Claude and Gemini are familiar examples. Businesses use it to draft content, answer customer chats and summarise documents, but a person must check the output: it can be confidently wrong.

This guide covers how it works, which tools exist, where it fits in a Thai SME, what can go wrong and how to start. For how generative AI differs from AI that carries out work on its own, read Generative AI vs Agentic AI.

What generative AI is: AI that creates, not just predicts

The AI most businesses already use, such as spam filters, product recommendations and credit scoring, judges existing data. Generative AI produces something new: a reply to a customer, a product image, a video script, a block of code.

NIST, the US standards institute, cites a definition of generative AI as the class of AI models that emulate the structure and characteristics of input data to generate derived synthetic content, including images, video, audio and text. In business terms, the model has seen enough examples to know what a good apology email looks like, and writes a new one on request.

How generative AI works, without the jargon

1. It learns patterns from huge amounts of data. Text-based generative AI runs on a large language model (LLM). Google's Machine Learning Crash Course explains that LLMs are trained on massive datasets by predicting hidden or upcoming pieces of text, again and again, until the model absorbs the patterns of language. What it keeps is not a database of answers but a statistical sense of what tends to follow what.

2. It answers one token at a time. The model doesn't look up a stored answer. It calculates the most suitable next token (a word, part of a word or a single character), adds it, and repeats. Most LLMs use the Transformer architecture, whose self-attention lets the model weigh how words relate. In "the customer was unhappy with the product because it arrived late", it works out that "it" is the product. Image generators often refine random noise step by step until the picture matches the description.

3. The prompt and context decide the quality. The prompt is your instruction; the context is what you attach, like a price list or returns policy. "Write a post selling sunscreen" gets something generic. "Write three Facebook posts selling sunscreen to outdoor workers, friendly tone, under 80 words, no medical claims, using the attached product sheet" gets something close to publishable.

The key point: it composes, it doesn't look things up. Because answers come from probabilities, the model can write fluent, convincing sentences that are wrong. This is hallucination. It also knows nothing about your prices, stock or warranty terms unless you supply them, which is why business AI systems connect to the company's real data.

Well-known generative AI tools

These off-the-shelf tools work on the web or in an app; OpenAI, Anthropic and Google also offer APIs for connecting their models to company systems. Prices are from the official pricing pages on 26 October 2026.

How Thai SMEs use generative AI

Marketing content. Facebook posts, TikTok captions, product descriptions and promotion emails are quick to draft. Keep a short brand-voice guide in every prompt, and have someone check prices and product claims before publishing.

Customer chat on LINE OA and Facebook. Shops that sell through chat answer the same questions daily: price, ordering, shipping, opening hours. An AI chatbot connected to a LINE Official Account through the Messaging API can answer from your product and promotion data, then hand over to staff for claims, complaints or anything outside its scope. iReadCustomer builds these assistants as part of its AI development service.

Document summaries. Meeting notes and long email threads get shorter, and invoice fields can go straight into a sheet. Contracts still need a responsible person to read the original.

Q&A over company documents (RAG). With retrieval-augmented generation, the system indexes your manuals, policies and specs, retrieves the passages relevant to each question, and tells the model to answer only from them and cite them. If nothing is found, it says so or routes the question to a person. Updating an answer means updating a document, not retraining a model. Plan access rights early so payroll or client contracts never surface for the wrong person.

Limits and risks

Confident errors. NIST calls this confabulation: confidently stated but false content that can mislead users, sometimes with invented reasoning or citations. Anthropic's guidance lists three basic defences: let the model say "I don't know", have it quote the source word for word before answering, and require a citation for every claim, retracting any it cannot support. In business systems, add a fourth: risky actions need human approval.

Personal data and PDPA. Thailand's Personal Data Protection Act B.E. 2562 (2019) covers data that identifies a person, such as names, phone numbers and purchase history. Don't paste customer data into personal AI accounts. Check training settings: Claude's pricing page lists its individual plans as opt-out for model training and says Team and Enterprise don't train on your content by default; Google says Gemini Enterprise does not use customer data, including prompts and outputs, to train its models. For custom systems, decide where data is stored, for how long and who can see it. Transfers of personal data abroad carry their own PDPA requirements, so involve your data protection officer or lawyer.

Copyright and trade secrets. NIST notes that generative AI makes it easier to reproduce copyrighted or trademarked content without permission and to expose trade secrets. Don't ask a tool to copy another brand's logo or characters, read each tool's commercial-use terms, and keep confidential formulas out of tools you haven't vetted.

Thai language and bias. NIST also flags performance gaps between groups and languages. Test with your customers' real Thai, including slang, typos and industry terms.

How to start

1. Off-the-shelf tools on internal work first. Drafts and summaries with no customer data, plus a one-page policy on what data is off-limits and what a person must check.
2. Build when the work demands it: staff copy data between your systems and an AI chat every day, the work faces customers, or you need a record of what the AI answered and from which data.
3. Pilot one task. Pick one frequent, measurable task, agree real test examples and a pass mark before building, and start with read-only access.
4. Measure with your own numbers. Record a baseline, then track time per task, questions resolved without hand-over, accuracy on your test set, AI cost per task, and errors that reached customers.

That is how iReadCustomer works: one task, tested on your real work before launch, routed to a person when the AI is unsure, then handed over to your team with code and documentation.

How much does generative AI cost?

There are three cost layers, and you may need only the first.

1. Tool subscriptions, per user per month (table above).
2. API usage, billed by the provider per token; rates vary by model and volume, so measure cost per task in the pilot. In systems iReadCustomer builds, the AI and cloud accounts are in your name and you pay the providers directly.
3. Custom development. iReadCustomer prices by the man-day (one engineer, one working day) from ฿7,990, and projects start at ฿50,000.

The figure depends on how many systems connect, your data and the cost of an AI mistake. See AI agent and chatbot pricing, or contact iReadCustomer to have your scope estimated.

Sources

• Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1) — NIST
• LLMs: What's a large language model? — Google Machine Learning Crash Course
• Reduce hallucinations — Claude Platform Docs (Anthropic)
• Pricing — Claude (Anthropic)
• Gemini plans — Google
• Personal Data Protection Act B.E. 2562 — Office of the Personal Data Protection Committee (PDPC)

General information as of 26 October 2026. Tool prices change, so check the official pricing pages before you decide. This is not legal advice.